Information Extraction from the Web by Matching Visual Presentation Patterns

The documents available in the World Wide Web contain large amounts of information presented in tables, lists or other visually regular structures. The published information is however usually not annotated explicitly or implicitly and its interpretation is left on a human reader. This makes the information extraction from web documents a challenging problem. Most existing approaches are based on a top-down approach that proceeds from the larger page regions to individual data records, which depends on different heuristics. We present an opposite bottom-up approach. We roughly identify the smallest data fields in the document and later, we refine this approximation by matching the discovered visual presentation patterns with the expected semantic structure of the extracted information. This approach allows to efficiently extract structured data from heterogeneous documents without any kind of additional annotations as we demonstrate experimentally on various application domains.